The announcement landed with the usual fanfare. Largan Precision, the Taiwanese lens giant that feeds Apple's camera modules, is partnering with TSMC on co-packaged optics. The market read it as a marriage of optical design and semiconductor muscle. The math is perfect; the reality is broken.
Between the commit and the block lies the trap. This is not a story about a new product. It is a story about a structural shift in how AI data centers will move data, and the uncomfortable truth that the two companies involved are entering this race with different clocks, different cultures, and different definitions of what constitutes a win.
I have spent the last four years dissecting semiconductor supply chains, and the CPO narrative has always smelled of over-optimism. The industry loves a good roadmap. The industry hates the yield curve. Largan and TSMC are about to learn the difference.
Context: The Hype Cycle Meets the Physical Layer
Co-packaged optics is the next logical step for AI data centers. The current architecture uses pluggable optical transceivers, which sit at the edge of the switch and convert electrical signals to light. They work, but they consume power, generate heat, and take up space. As AI clusters scale to tens of thousands of GPUs, the interconnect becomes the bottleneck. CPO moves the optical engine onto the same substrate as the switch or compute chip, eliminating the electrical-to-optical conversion at the package edge. The promise is lower latency, lower power, and higher bandwidth density.
TSMC has been pushing this agenda for years. At its 2024 North America Technology Symposium, the company showcased its COUPE platform, a compact universal photonic engine slated for 2025. Largan's role is to design the optical engines, the lenses, and the coupling mechanisms that make the photonic link work. On paper, it is a perfect match. TSMC owns the packaging, Largan owns the optics.
But the paper is where the problems begin. CPO is not a single technology. It is a stack of technologies, each with its own failure modes. Silicon photonics, micro-ring modulators, fiber coupling, thermal management, and test methodologies. Every one of these is a discipline in itself. Largan has spent decades perfecting plastic and glass lenses for smartphones. That expertise does not transfer cleanly to the world of sub-micron optical alignment on a semiconductor substrate.
Core: The Systematic Teardown
Let me be precise about what this partnership actually entails, and where the value will be created or destroyed.
The Yield Problem
The first and most critical issue is yield. TSMC's CoWoS packaging line has matured to over 90% yield. That is the benchmark. CPO introduces new processes: optical coupling, laser integration, and the alignment of photonic components with electronic ones. The tolerance for misalignment is measured in nanometers. A single dust particle can kill a package. The industry is still in the early stages of ramping CPO yield, and the numbers are not pretty. If Largan's optical engines cannot hit a yield above 90%, the cost structure of the entire CPO solution collapses.
I have audited enough semiconductor fabs to know that yield is not a linear function of effort. It is a function of process maturity, and process maturity takes time. The expectation that CPO will reach mass production in 2025-2026 is optimistic. The expectation that it will do so at a cost competitive with pluggable optics is fantasy.
The Cost Structure
The value chain for CPO is different from traditional optics. The optical engine accounts for 30-40% of the module cost, and packaging accounts for 40-50%. This is a fundamental shift. In the traditional pluggable market, the optical module maker controls the entire assembly. In the CPO world, the packaging is controlled by TSMC, and the optical engine is controlled by Largan. This creates a two-vendor dependency that will complicate procurement and pricing.
Largan's gross margins have historically been 60-65%, driven by its near-monopoly in high-end smartphone lenses. The company's margins have been eroding, from over 70% in 2019 to around 60% in 2024, as competition in the smartphone lens market intensifies. The CPO business is expected to have gross margins of 60-70%, which would be a welcome boost. But that assumes the yield problem is solved. If yield is below 90%, the effective cost per good unit will be significantly higher, and the margin will be compressed.
The Capital Expenditure Trap
Largan's capital expenditure has historically been 10-15% of revenue. The CPO business will require new production lines for optical engines, which means a significant increase in capex intensity. This will pressure free cash flow in the short term. The company's operating cash flow is around $500-600 million per year, and free cash flow is around $300-400 million. A major capex push for CPO will eat into that.
TSMC, on the other hand, spends 35-40% of revenue on capex. The CPO line is a rounding error for them. But the strategic importance is not financial. It is about maintaining dominance in advanced packaging. TSMC's CoWoS capacity is already sold out, driven by NVIDIA's AI GPU demand. Adding CPO capacity will require new fabs and new equipment, with a lead time of 12-18 months. The depreciation burden will be significant, and I estimate it will drag gross margins by 2-3 percentage points in the initial phase.
The Market Reality
The market for CPO is real, but the timing is uncertain. LightCounting projects the CPO market will grow from $500 million in 2024 to $5 billion in 2028, a compound annual growth rate of 60%. That is a compelling number. But it is a projection, not a guarantee. The adoption of CPO depends on the hyperscalers, and they are notoriously conservative when it comes to changing their data center architectures. The pluggable optics market is not going to disappear overnight. Companies like InnoLight and Eoptolink have a 2-3 year window before CPO becomes a real threat.
NVIDIA's GB200 platform, expected to ship in 2025, will require CPO optical engines. But NVIDIA is also a master of supply chain management, and it will not bet its entire roadmap on a single technology that is still in the yield ramp phase. The demand is there, but the supply is not ready.
The Competitive Landscape
Largan and TSMC are not alone in this race. Intel has been investing in silicon photonics for over a decade, with an annual R&D budget of around $1 billion. Broadcom has its own CPO switch chips and is a key partner for TSMC. Marvell is developing CPO DSPs. The combined Largan-TSMC effort is in the first tier, but it is not a monopoly. The technology is still in flux, and the competitive landscape will not be settled for another 2-3 years.
Largan's customer concentration is a major risk. Apple accounts for over 50% of its revenue. The CPO partnership is a clear attempt to diversify, but it will take years to build a meaningful revenue stream. The company's R&D spending is around $200-300 million per year, which is small compared to Intel's $1 billion or Broadcom's $500 million. Largan is betting that its optical design expertise will be enough to compete. That is a risky bet.
Contrarian: What the Bulls Got Right
I have been harsh, but the bulls are not entirely wrong. The logic of CPO is sound. The power and latency advantages are real. The demand from AI data centers is not a fad. And the combination of Largan's optical design and TSMC's packaging is genuinely complementary. There is no other company in the world that has both the optical expertise and the semiconductor packaging capability to execute this at scale.
TSMC's COUPE platform is a serious piece of engineering. The company has a track record of solving hard problems, from 3nm process technology to CoWoS packaging. If anyone can make CPO work, it is TSMC. And Largan's lens design capabilities are world-class. The company has been making precision optics for decades, and it has the intellectual property to back it up.
The strategic logic is also sound. Largan needs a second growth curve. The smartphone market is saturated, and the company's dependence on Apple is a structural weakness. CPO offers a path to diversification. TSMC needs to extend its advanced packaging moat. CPO is the natural next step. The partnership makes sense on every level except the one that matters most: execution.
Takeaway: The Accountability Call
The math is perfect; the reality is broken. The CPO partnership between Largan and TSMC is a bet on the future, but the future is not a straight line. The yield problem is the elephant in the room. If Largan cannot deliver optical engines with yields above 90%, the entire project will be delayed, and the cost structure will be untenable. The market is pricing in a 2025-2026 mass production timeline. I am skeptical.
Logic holds; incentives collapse. Largan's incentive is to diversify away from Apple. TSMC's incentive is to maintain its packaging dominance. But the incentives of the engineers on the ground are different. They are measured on yield, on cost, and on time-to-market. The pressure to ship will be immense, and the temptation to cut corners will be real.
Trust is a variable that must be zero. I do not trust the roadmap. I do not trust the projections. I trust the data. And the data says that CPO is still in the early stages of the yield ramp. The next 12-18 months will be critical. If Largan and TSMC can hit their milestones, the CPO market will be theirs. If they slip, the window will close, and Intel or Broadcom will be waiting.
The illusion breaks when the liquidity dries up. In this case, the liquidity is not money. It is time. And time is the one resource that cannot be bought. The question is not whether CPO will work. It is whether Largan and TSMC can make it work before the market moves on to the next shiny object. Every transaction is a potential extraction point, and in this case, the extraction point is the yield curve. Watch it closely.
